How marketers are measuring AI search

Marketing has always had a measurement gap. Dark social, word-of-mouth, a podcast mention that sparks a wave of branded search—none of these show up cleanly in a dashboard. Teams have always learned to adapt with incomplete data, fill in the gaps with proxies, and make do. Generative Engine Optimization (GEO) attribution feels familiar, but with critical updates impacting strategy: 

• The pre-click problem: When someone uses AI to research a category, evaluate brands, and narrow their options, they arrive at your site ready to convert with an invisible research journey behind them. By the time they show up in your analytics, the decision is largely made. Last-click attribution undercounts GEO’s contribution and misrepresents the entire funnel. The click you’re crediting could be the confirmation, not the conversion.
• The traffic paradox: The instinct when organic traffic drops is to treat it as a performance problem by compensating with paid, producing more content, and filling gaps in the funnel. But new data complicates that conventional wisdom. For Paula at Wix, non-branded traffic fell while average conversion per user grew by double digits year over year. AI-driven traffic is tighter, but it sends you stronger prospects. Optimizing for traffic volume in that environment can obscure what’s actually working: that your content is reaching your niche, target audience.

These changing dynamics mean that the standard attribution playbook produces the wrong conclusions when applied to GEO. The new method for many marketers in today’s landscape is to build their own measurement stack.

While perfect GEO attribution doesn't exist, marketing leaders across our portfolio are far from idle. We sat down with Natalia Bandach, Senior Director of Growth Marketing at Cloudinary and Paula Ximena Mejia, VP of Marketing at Wix, among other marketing leaders across vertical SaaS and AI teams, to learn what they’re seeing across their workflows, what they’re doing as first movers, and what’s yielding results. 

Key insights for marketing leaders tackling the AI attribution problem

• Stop waiting for perfect GEO attribution. Start triangulating. AI assistants don’t pass referral data the same way traditional search does. The goal isn’t to fix it but to build a proxy system that gives you enough signal to make decisions and improve.

• Less traffic can mean better business outcomes. Non-branded search is down. GEO-sourced users arrive having already done their research with AI. They know what brands they want before ever clicking on a SERP result.
• Last click was always a convention. Last-click attribution wasn’t an accurate reflection of how decisions were made; it was just the most measurable proxy available. GEO makes it hard to maintain the status quo, but the underlying problem isn’t new.
• GEO requires company-wide buy-in. GEO touches content, social, PR, community, and even employee-generated content. The companies making the most progress have established KPIs across every team involved in branded assets and communications and made GEO a company-wide priority, not a marketing project.
• Start measuring imperfectly now. The teams systematically underestimating GEO’s contribution are the ones waiting for clean attribution before they commit. Get your proxy stack in place, review weekly, and treat every measurement decision as an experiment you’ll refine.

Stop waiting for perfect GEO attribution. Start triangulating.

AI assistants don’t pass the same kind of referral data that traditional search touchpoints did, so marketers are now forced to build proxy systems that guide them in the right direction. Natalia at Cloudinary stopped looking for a single metric to measure AI and built a system of three proxies that, together, give her enough signal to move with more accuracy and confidence. “Whoever says they have it perfectly set up is probably trying to sell something,” she shared. “It doesn’t really exist right now.”

1. Her first proxy is an AI visibility score, tracked through Profound top-down and through Google Search Console bottom-up. Cloudinary targets a 75% overall share of voice on its tracked keyword prompt groups in Profound. These are specific questions and keyword clusters where they want to show up in AI-generated answers. Natalia’s benchmark for what good looks like: Cloudinary currently holds a 91% AI visibility score on the question “What's a reliable image API for automatic pre-processing?”—a direct reflection of the product-led authority they’ve built through years of developer adoption and embedded-product growth. The number came from being genuinely embedded in the category, not a GEO campaign.
2. The second proxy is self-reported attribution, or the classic “How did you hear about us?” survey. Natalia’s operational detail here is easy to overlook but worth noting: shuffle the response options periodically to avoid anchoring bias (i.e., occasionally, make the AI referral the first option). She also recommends a validity check where you split-test questions between two different audience cohorts and look for consistencies. If things aren’t holding up across cohorts, something is off in how the data is either being collected or read.
3. Her third proxy is referral traffic from AI chatbots. In GA4, she tracks ChatGPT, Perplexity, and Claude as identifiable sources, now also surfaced under ‘AI Assistant.’ This is the most familiar measurement surface for most teams, and the most incomplete on its own because it only captures users who clicked from an AI interface, not the larger group who were influenced by an AI answer and then arrived through a different path.

A vertical software CMO approached the same problem from a different direction. Where Natalia builds a measurement stack to capture what analytics miss, this CMO reads the anomalies already in their analytics and assigns reasonable meaning to them. “Don’t let the loss of attribution fidelity stop you from being creative,” they advise. 

When an “unknown” source spiked in their data one quarter, their team pulled it out of the unknown, attributed it to AI, and tracked accordingly. The logic is that if someone lands on a deep, specific product page that was getting no traction before, there’s only one way they found it. “Would I say it’s a 100% science? Absolutely not. But that’s the nature of the beast right now.” 

Natalia’s proxy system and the CMO’s behavioral heuristics are solving the same problem from different directions, but both capture signal in place of a single clean metric.

Less traffic can mean better business outcomes

Non-branded search is down across the board. The instinct is to treat that as a performance problem by compensating with paid, producing more content, and filling the gaps. That instinct is usually wrong when it comes to optimizing for AI search.

At Wix, Paula noticed non-branded traffic fell while average conversion per user grew double digits year-over-year. Another team saw conversion to pipeline from GEO-sourced leads run at triple their channel average. “I’ve never in my career seen a channel increase pipeline 3x,” that CMO shared.

The explanation is a structural one. GEO users arrive after doing their research through AI. They've evaluated options, compared alternatives, and formed a preference before they ever visit your site. The funnel has narrowed at the top and become more efficient in the middle. AI-driven traffic is smaller, but the buyers it sends are better.

Paula flags the risk of teams that don’t recognize this shift and reach for paid to prop volume numbers back up. “You can end up artificially propping traffic for the sake of not losing traffic," she says. “Traffic isn't the goal—conversions are.” According to Paula, the more useful question for CMOs right now is what is the actual relationship between your organic and paid spend? Is the incremental paid investment buying real conversion, or just covering a number on a dashboard?

Last click was always a convention

Last-click attribution was never an accurate reflection of how decisions were made, just the most measurable proxy available. GEO makes the fiction harder to maintain, but the underlying problem isn't new. “Decisions now start to happen before the click,” says Paula. “Educational content that used to drive traffic may still be shaping decisions—we just can't see it anymore.”

The implication is that the conversion marketing teams have long relied on was always a simplification, and GEO has made that simplification untenable. An educational resource that shaped a buyer's decision inside an AI conversation, before they ever visit your site, won't show up anywhere in your analytics. 

A more resilient indicator, Paula argues, is how your branded search is doing. People searching your brand name signals that demand is being created somewhere upstream—through AI recommendations, editorial coverage, community mentions—even when you can't trace exactly where. It's the footprint of influence that doesn't require a click to exist, and it's far more durable than non-branded traffic in a world where AI is increasingly resolving queries before someone reaches your site.

“Our organic growth teams used to spend a lot of time creating a lot of content,” says Paula. “I now ask myself: Is all content worth creating? The answer is no. Saying yes to something comes at the cost of saying no to something else.”

The goal now is to address what some call your brand’s “language layer” gap, where instead of pushing even more content, you optimize content to align with your brand ethos. This means ensuring your content shares a “common language” across all of your communications—and that third-party language reflects the same messaging.

GEO requires company-wide buy-in

A single team can own SEO because they control most of the inputs, manage the outputs, and can mostly silo the work from the rest of an organization. GEO can't be contained that way as the fundamentals span content, PR, community, social, and even employee-generated content. 

At Wix, this meant establishing what Paula references as “semantic triples” or a common language across every team that creates assets. Whether it's an SEO page, a community post, or a journalist briefing, every piece of content describes Wix's products using the same specific language, reinforcing the same brand-topic associations in the models. The biggest unlock from this approach has been through PR. At Wix, PR sits outside marketing, but now journalists are briefed using the same vocabulary before every major story. “If it gets published, it helps,” Paula says. “It's become almost an operational requirement for GEO to be effective.”

At Cloudinary, Natalia says they structure GEO as a flywheel divided between branded and non-branded, with distinct owners at each stage of the funnel. Each owner runs their own queries and experiments. Their newest motion is what she calls AX, or agentic experience: designing content and product surfaces for AI agents making decisions on behalf of users. “That's critical for infrastructure companies,” she says. 

A CMO at a vertical AI company has a more cultural approach. Every marketing meeting opens with ten minutes of “who's done something new with AI?” The company’s last offsite was dedicated entirely to GEO. The directive to take GEO seriously, they say, has to come from the CEO down: “It has to be non-negotiable. This is such a transformational moment.”

Start measuring AI imperfectly now

The teams systematically underestimating GEO's contribution are the ones waiting for clean attribution before they commit. The ones gaining an advantage are those who have already started measuring anyway. 

The vertical AI CMO adds a tax to their self-reported survey data whenever response rates are low enough that raw numbers undercount meaningfully, inflating the numbers to help draw more substantial conclusions. They also tag behavioral anomalies as GEO based on heuristics because their validation isn't methodological purity. “We’re systematically underestimating what we're allocating to GEO,” they say. 

Their team also runs a weekly review cycle on visibility score, sentiment and truth score, and recommendation position—an imperfect yet structured way to stay close enough to a fast-moving channel to catch shifts as they happen. “Things move too fast now for quarterly reviews,” says the CMO. 

Natalia says they run 120–150 experiments per quarter at Cloudinary and are scaling up for GEO. Each manipulation for AI is treated as an experiment. One experiment they conduct is the use of recency keywords, or keywords that are time-sensitive. “Every model has a knowledge cutoff, and recency keywords—like ‘best XYZ in 2026’—trigger a live search for the latest information. This is a fast win when you're trying to be visible within a week or two on specific queries.” According to Natalia, content optimized with recency keywords can show up in 2-4 weeks.

The common thread is a willingness to experiment and act on directionally accurate data rather than wait for precise data. Build the proxy stack, review regularly (weekly), and treat every measurement call as an opportunity to iterate as the landscape matures.

Build your own GEO measurement stack

These learnings should serve as a tentative, but actionable guide to GEO measurement. Altogether, these marketing leaders’ approaches can help you better understand your brand’s position in AI search. A proxy system gives you tracking infrastructure, behavioral heuristics are a way to read the data you already have, and an attribution reframe tells you what you’re actually optimizing for. 

To build your own AI measurement system, start by tracking these metrics to refine or experiment with:

KPI	Discovery question	Method
AI visibility score	Is AI finding and citing us?	Profound or similar AI search tool; track brand visibility and presence across topics and platforms
AI referral traffic	Which users clicked through from AI?	GA4 source/medium or similar
Behavioral anomalies	Which sessions could have been AI-influenced?	Flag deep product page spikes; pull any ‘unknown’ traffic and tag as GEO
Self-reported attribution	What are users telling us directly?	Shuffle survey options and 3x your self-reported data when responses are low
Branded search volume	Are we building durable brand demand?	Google Search Console or similar
Pipeline conversion rate	Are GEO users converting?	Compare AI-sourced sessions against channel average